Original URL: http://www.theregister.co.uk/2007/12/11/hp_gpgpu_fpga/

HP plots path to server accelerator madness in '08

Radical conservatism

By Ashlee Vance

Posted in Servers, 11th December 2007 12:54 GMT

Let's be clear right up front. HP refuses to confirm or deny much of anything in this story. But, if we read the non-answers provided by HP high performance computing (HPC) manager Ed Turkel right, then HP has a heck of a lot of goodies prepared for its high-end server customers next year.

In our estimation, HP has led all of the Tier 1 server vendors on the server accelerator front. It has backed companies working on co-processors and those creating tools used to write improved multi-threaded software with marketing dollars and engineering aid. In addition, HP has crafted a number of HPC servers for cluster and blade set-ups, helping it take a solid second place after IBM in the supercomputer race. And, unlike IBM, HP has emphasized a very broad set of acceleration tools rather than spending most if its time and energy hawking a single homegrown product like the Cell chip.

Come 2008, HP will build on these efforts – we think – by adding things such as GPGPUs (general purpose GPUs) and Opteron socket-ready FPGAs (field programmable gate arrays) to its price list.

With regard to GPGPUs, for example, Turkel said, "That's the area where we would resell something first."

Turkel emphasized the use of the word "would" because he's in no position to pre-announce products or initiatives – wink, wink – but in the land of the hypothetical, HP finds GPGPUs pretty interesting.

The vendor has fired up talks with AMD/ATI and Nvidia around its Tesla systems to see what makes sense from a resale perspective. Customers can, for example, take a single GPGPU card and slot it into existing systems in order to speed up certain types of software that runs better on the graphics chips' numerous engines than on dual- or quad-core x86 chips. Or you can go the Tesla route and buy a deskside or rack unit with numerous GPUs installed to get a super-charged speed-up.

(Anyone who wants to be brought up to speed on Tesla or the other accelerators out there can start here.)

According to Turkel, HP would probably just do the GPGPU cards at first, since they're so easy to plug into workstations and servers. Tesla, however, seems to be getting a ton of attention and could well make its way onto HP's price list. If so, that would be a major win for Nvidia.

With regard to FPGAs, a pair of start-ups – DRC and XtremeData – currently sell units that plug into Opteron processor sockets. Companies in the oil and gas, life sciences and financial services fields have been tweaking their software to let specific routines run on the FPGAs. As a result, they've seen huge speed-ups in performance – say, 10x to 40x – and much better performance per watt figures. If you're willing to do the software work and have talented coders on retainer, the FPGA play can make sense.

Down the road, DRC and XtremeData will have similar FPGAs for Xeon-based systems. (Update: XtremeData has already come to market with a product that goes into Xeon sockets, and SGI is using it.)

"We're talking to both of them," Turkel said. "I think they both have an interesting approach with the socket-based solutions. We probably will end up doing both."

And from what we hear from other sources, that probably is a definite yes.

Accelerators currently dominate a lot of the HPC discussions, although there's another trend in this part of that marker that has captured HP's curiosity – mega data centers.

The likes of Google, Microsoft, Yahoo! And Amazon don't meet the classic HPC model, but they certainly seem closer to national labs than corporate types with their $500m-$600m data centers packed with thousands of servers and storage systems.

With that in mind, HP is engaging in internal discussions about how to handle mega data center types.

The company has discussed creating specialized low-power servers with under-achieving memory and cheap disks in a bid to cater to the mega data center set. Dell does something similar today for large orders through a custom data center program. In addition, Rackable Systems has met the needs of folks such as Microsoft and Amazon through its low-power systems.

As it turns out, a lot of Rackable's historical business has come through a reselling-type partnership with HP. Now, HP may go ahead and lead these efforts on its own.

"I don't want to pre-announce anything," Turkel said. "We are certainly looking at those markets and how we might be able to satisfy them."

While HP won't preannounce much of anything, it will talk about existing HPC wins.

We know. You're shocked.

The US, Western Europe and Japan have treated HP well to date in the supercomputer arena and now HP wants to tap into the BRIC (Brazil, Russia, India and China) nations with its fanciest gear.

HP has expressed its BRIC enthusiasm by pointing us – and consequently you – to a new list of China's top supercomputers. As it turns out, HP dominates the list, although IBM, as usual, has the top box.

HP's overall HPC play marks a continued trend to "balance" R&D work with partnerships.

You'll see IBM and Sun Microsystems at one of the spectrum building things such as their own chips, switches and file systems. And then Dell slots in at the other end of the spectrum by partnering in most cases and taking a wait and see approach to which bits and bobs take off.

HP proves more willing than Dell to stomach risk in the HPC game and is aggressive about partnering with start-ups, since this is the only way a company unwilling to invest as much R&D money as in the past can stay on the cutting-edge.

It's a sort of radical conservatism that appears to be serving HP will to date. ®